CELF Project Proposal/Add Device Tree emulation support to QEMU

From eLinux.org
Revision as of 04:12, 8 January 2010 by Jw petalogix (Talk | contribs) (Comments)

Jump to: navigation, search
Summary 
Add Device Tree emulation support to QEMU


Proposer 
Rob Landley

Description

This task would be to integrate a flattened device tree parser into the emulator QEMU, so that QEMU can create board emulations on the fly (at runtime) from the same data files the Linux kernel uses to attach drivers to hardware.

Currently, the QEMU is emulating system boards via hardwired .c files, which explicitly set up the resources for each emulation in a separate C function. In theory, QEMU could parse the same device tree data format the Linux kernel uses to set up its hardware resources, and then pass hardware resources along to kernels it invokes through its built-in bootloader (I.E. with the -kernel option). This could allow new boards to be added to qemu simply by supplying device tree files at runtime (assuming emulations for the appropriate peripherals had already been implemented in QEMU).

Device Tree information

The Device Tree format is a reasonably generic data file describing hardware layout. It is documented in the linux kernel source at:

 Documentation/powerpc/booting-without-of.txt

It started as the data structure Open Firmware used to describe supported hardware to operating systems, and was picked up by bootloaders such as u- boot. It allows the Linux kernel to parse a generic data structure at boot time to configure itself for the current hardware layout, instead of hardwiring board support in individual .c files.

Device trees are created using an ascii format to describe a board layout, which is converted into a flattened binary representation by dtc (the "device tree compiler", included in current linux kernel sources in scripts/dtc). A bootloader supporting device trees loads the flattened device tree into memory as a binary blob, and passes the linux kernel a pointer to this blob in a register. The kernel then uses a built-in device tree parser to understand the board's hardware layout and initialize itself. (Depending on kernel .config, this information may also be queried from userspace via a /proc interface.)

At the device tree BOF at OLS in 2008, a number of developers expressed interest in extending device tree support to other architectures (such as arm, mips, and sh4). As a result, device tree development was moved off of the PowerPC mailing list to its own list, for the purpose of genericizing it to more architectures:

 https://lists.ozlabs.org/listinfo/devicetree-discuss

The Linux MAINTAINERS entry for device tree support is:

 OPEN FIRMWARE AND FLATTENED DEVICE TREE
 M:      Grant Likely <grant.likely@secretlab.ca>
 L:      devicetree-discuss@lists.ozlabs.org
 W:      http://fdt.secretlab.ca
 S:      Maintained
 F:      drivers/of
 F:      include/linux/of*.h
 K:      of_get_property

In theory, at some future date the kernel will no longer need hardwired .c files describing the layout of boards for any of these architectures. Instead it can have a device tree for each board, which can be statically linked into the kernel binary if necessary.

If the kernel doesn't need board support hardwared in .c files, there's no reason QEMU needs it either.

Related work

Scope 
Unknown

Comments

Anthony Liguori writes:

See http://thread.gmane.org/gmane.comp.emulators.qemu/44869

I'm not sure why Paul never pushed it but I think he was able to create the syborg board purely from a device tree.


Blue Swirl writes:

In addition to -kernel case, the device tree should be passed to OpenBIOS, which would generate OF tree from DT. This would be useful for OSes other than Linux.

Another interesting case is sun4v hypervisor, which uses a similar device tree. Some DT transformation would be required in QEMU.


John Williams writes:

We've made a pretty good start on this for the MicroBlaze architecture. Our kernel platform config is driven by device trees, and we can now create QEMU systems dynamically, straight from the same device tree. There are no doubt many things that can be improved, but to have a working starting reference will be useful I'm sure.